EduInsights

View saved version →

---
name: audit-eduinsights-evidence
description: Audit an EduInsights answer, comparison, SQL result, or research method for source provenance, release and period alignment, entity resolution, level of detail, coverage, applicability, additive measures, and unsupported causal claims. Use when the person asks to verify, review, reproduce, challenge, or quality-check education or workforce evidence.
---

# Audit EduInsights evidence

Test whether each material claim is supported by the returned evidence. Do not rerun all research automatically; inspect the existing trace first and use tools only to fill a specific audit gap.

## Audit the claims

Read [audit-checklist.md](references/audit-checklist.md). Apply every section that matches the answer.

Build a compact claim ledger with:

- the claim;
- the measure and unit;
- the entity and level of detail;
- the period and release;
- the source;
- the evidence fit;
- the coverage reason;
- the audit result.

Use `edu_get_sources` for exact release details when the trace contains Source Release IDs. Use `edu_describe_data` for relation meaning. Use `edu_run_sql` only when a missing check requires a new query and focused tools cannot answer it.

## Rate findings

Use these results:

- **Supported:** the claim matches the source, period, entity, measure, and level of detail.
- **Supported with a limit:** the evidence answers the main question, but its population, geography, or level is narrower or broader.
- **Not established:** the evidence is relevant but does not measure the claimed outcome or relationship.
- **Cannot verify:** the trace lacks the needed source, release, period, or result rows.

Do not call a bounded source invalid merely because it does not cover every product or population. Assess whether it measures the concept asked about and state its scope.

## Report the audit

Lead with the most consequential issue. Separate correctness errors from disclosure improvements. Give the smallest concrete fix for each failed check.

Finish with a short list of claims that remain safe to use and claims that should be revised or removed.

Referenced files: 2

View saved version →

---
name: draft-eduinsights-brief
description: Turn EduInsights college, field, career, workforce, accreditation, or AI-use research into a concise decision brief for a provost, dean, workforce leader, policymaker, or advisor. Use when the person asks for a briefing, memo, recommendation, comparison summary, or decision-ready synthesis rather than a raw data answer.
---

# Draft an EduInsights brief

Build the brief from verified evidence. If the provided research lacks a material measure, source, period, or comparison, use the EduInsights tools to fill that gap before drafting.

## Set the decision

Identify:

- who will use the brief;
- the decision they need to make;
- the options or entities under consideration;
- the period and geography;
- the evidence that would change the decision.

Ask a question only when the missing choice would materially change the recommendation.

## Research proportionately

Use `edu_resolve_entity` for names. Prefer `edu_get_institution`, `edu_get_program`, `edu_get_occupation`, `edu_compare`, and `edu_aggregate`. Use `edu_get_sources` to verify releases. Use SQL only for a necessary question the focused tools cannot answer.

Do not fill gaps with assumptions. Mark an important unanswered part as an evidence gap and explain what source would resolve it.

## Draft the brief

Read [brief-template.md](references/brief-template.md). Use only the sections that help the decision.

Lead with the decision and the strongest supporting finding. Keep findings separate from judgment. Tie every figure to a year and source. Translate internal data terms into ordinary language.

Do not show codes, relation names, or SQL unless the audience needs a reproducible appendix.

## Check the result

Before finishing:

1. confirm the recommendation follows from the cited evidence;
2. confirm comparisons use aligned periods and measures;
3. state the strongest alternative explanation or limit;
4. distinguish reported degrees from current catalog programs;
5. distinguish related careers from graduate destinations;
6. distinguish observed AI use from employment forecasts;
7. make the next action specific and proportionate to the evidence.

Referenced files: 2

View saved version →

---
name: research-with-eduinsights
description: Research U.S. colleges, reported fields and credentials, enrollment, completions, outcomes, accreditation, careers, occupational tasks, pay, outlook, and observed AI use with the EduInsights MCP tools. Use for factual questions, rankings, comparisons, and program-to-career questions that need public evidence. Also use when deciding which EduInsights tool should answer a question. Do not claim current catalog courses or requirements from this data.
---

# Research with EduInsights

Answer the person's question with the smallest suitable read-only tool path. Keep every claim connected to its source, period, and level of detail.

## Start with the question

Identify these parts before calling a tool:

- the named college, field, credential, occupation, task, or place;
- the measure the person means;
- the requested period or whether current accepted evidence is enough;
- whether the task is a focused lookup, comparison, grouped result, history, or source check.

If a named entity is ambiguous, resolve it. Ask the person only when ranked candidates remain genuinely ambiguous.

## Route the work

Read [tool-routing.md](references/tool-routing.md) before the first tool call in a conversation. Follow its focused-tool-first route.

For a cross-domain question or any question about what the evidence can establish, also read [question-to-evidence.md](references/question-to-evidence.md).

For SQL or a complex aggregation not covered by a focused tool, read [query-recipes.md](references/query-recipes.md) before calling `edu_run_sql`.

## Preserve meaning

- Keep the person's ordinary concept as the target. Do not rename an AI question into a Claude-only question.
- When Anthropic is the only relevant empirical source, use it as the best available AI-use evidence. State the product, population, period, and release.
- Treat reported degrees as historical reporting evidence, not proof of a currently offered named program.
- Treat field-to-career links as preparation paths, not observed graduate destinations.
- Treat observed AI use as behavior evidence, not a forecast of job loss, productivity, or wages.
- Treat broader field or national evidence as context for a narrower question. Label that broader fit instead of presenting it as direct evidence.
- Never turn a blank into zero. Use the supplied coverage reason.

## Verify before answering

Check that:

1. every name was resolved to a canonical identifier before filtering;
2. every figure retains its period and source or release;
3. compared figures use compatible measures, populations, places, and periods;
4. sums use additive measures only;
5. medians, wages, rates, scores, years, and overlapping source packages were not summed;
6. the answer states important boundaries once, close to the affected claim.

## Write the answer

Read [answer-contract.md](references/answer-contract.md) before producing the final response.

Lead with the answer. Use readable names rather than internal relation names. Show identifiers only when requested or needed for disambiguation. Pair every shown identifier with its label.

Finish only when the result includes the finding, the relevant year or period, the source scope, and the limit that would change a decision.

Referenced files: 5

View saved version →

---
name: understand-eduinsights-ontology
description: Explain and apply the EduInsights semantic world model, including canonical entities, identifiers, relationships, native grains, applicability, coverage, source boundaries, and the most important marts. Use for questions about the EduInsights ontology, semantic layer, data model, schema, relationship map, mart meanings, or which mart fits an education, workforce, accreditation, outcome, or AI-use question.
---

# Understand the EduInsights ontology

Explain how EduInsights represents the person's question before selecting data. Preserve the question's ordinary meaning while keeping every connection at its supported level of detail.

## Follow this path

1. Identify the requested entity, measure or outcome, population, place, and period.
2. Read [world-model.md](references/world-model.md) before explaining an entity, relationship, identifier, applicability rule, or coverage state.
3. Read [important-marts.md](references/important-marts.md) before naming, comparing, or selecting a `marts.*` relation.
4. When current availability or columns matter, call `edu_describe_data` without a topic. Then inspect the selected relation with `topic: table:marts.<relation>`.
5. Use the smallest relation at the question's native level of detail. Move across entities only through an explicit identifier or relationship.
6. State the connection type and the boundary that prevents overinterpretation.

Treat live `edu_describe_data` output as the authority for deployed availability and fields. Its overview can emphasize common starting views rather than print every relation. Treat a curated mart as unavailable only when its direct table description fails. Never query `marts.semantic_catalog` directly.

## Keep the two registers separate

Use exact entity, relationship, identifier, and mart names for builders, data work, and reproducibility. For everyone else, explain the plain meaning first. Add the technical name only when it helps the person inspect or reproduce the result.

For example, say "degrees a college reported in a field and year" before `ProgramAwardCell`. Say why a value is blank instead of exposing a coverage-state code.

## Protect the meaning

- Treat reported awards as historical reporting, not proof of a current catalog program.
- Treat broader field outcomes as context for a narrower field or program.
- Treat CIP-to-SOC links as preparation relationships, not observed graduate destinations.
- Keep occupations, industries, tasks, and skills distinct.
- Treat observed AI use as behavior in the named product and population, not employment impact or universal capability.
- Preserve source release, native level of detail, applicability, coverage, and identifier uncertainty.
- Keep zero distinct from suppressed, uncollected, inapplicable, nonparticipating, stale, or unresolved data.

When the person asks for an empirical result rather than an explanation, apply these semantic rules with `$research-with-eduinsights`.

## Finish with a usable map

Complete the answer only when it identifies:

- the person's concept and the matching ontology entity or entities;
- the relationship path and whether each step is direct, inherited, structural, empirical, or contextual;
- the best current mart or focused tool and what one row represents;
- the source, period, coverage, and limitation most likely to change the interpretation.

Referenced files: 3